This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2019.2911602, IEEE
Transactions on Neural Systems and Rehabilitation Engineering
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <

A Patient-Specific Single Sensor IoT-Based

Wearable Fall Prediction and Detection System
Wala Saadeh, Member, IEEE, Saad Adnan Butt, Student Member, and Muhammad Awais Bin Altaf,
Member, IEEE

Abstract— Falls in older adults are a major cause of morbidity

and mortality and are a key class of preventable injuries. This
paper presents a patient-specific (PS) fall prediction and detection
prototype system that utilizes a single tri-axial accelerometer
attached to the patient’s thigh to distinguish between activities of
daily living (ADL) and fall events. The proposed system consists of
two modes of operation: 1) fast mode for fall predication (FMFP)
predicting a fall event (300msec-700msec) before occurring, 2)
slow mode for fall detection (SMFD) with a 1-sec latency for
detecting a fall event. The nonlinear Support Vector Machine
Classifier (NLSVM)-based FMFP algorithm extracts 7
discriminating features for the pre-fall case to identify a fall risk
event and alarm the patient. The proposed SMFD algorithm
utilizes a Three-cascaded 1-sec sliding frames classification
architecture with a linear regression-based offline training to
identify a single and optimal threshold for each patient. Fall
incidence will trigger an alarming notice to the concern healthcare
providers via the internet. Experiments are performed with 20
different subjects (age above 65 years) and a total number of 100
associated falls and ADL recordings indoors and outdoors. The
accuracy of the proposed algorithms is furthermore validated via
MobiFall Dataset. FMFP achieves sensitivity and specificity of
97.8% and 99.1%, respectively, while SMFD achieves sensitivity
and specificity of 98.6% and 99.3%, respectively, for a total
number of 600 measured falls and ADL cases from 77 subjects.
Index Terms— Fall detection, fall prediction, feature extraction,
gait monitoring, patient-specific, support-vector machine,
threshold detection, wearable sensor.
I. INTRODUCTION
The rising populace of elderly people has increased the risk of
accidental and unassisted fall events [1]. Accidental falls are a
major concern for the elder people; being the main cause for
hospitalization and the second leading cause of unintended
injury-related demises among the elder people in the world [1]-
[3]. According to the U.S. Centers for Disease Control and
Prevention, 25% of the Americans aged 65 or above experience
a fall once each year [4]. Every 11 second an elderly person is
treated in the hospital for a fall and, every 19 seconds an elderly
person dies from a fall. These falls cause more than 27,000
deaths and more than 800,000 injuries annually and cost tens of
thousands million dollars every year [1], [2]. Fig. 1 shows the
number of reported fatal fall events and their expenses per age
group for elder people (> 65 years) in the USA only in 2010 [3].
Fig. 1: Distribution of fatal fall events and cost per age group in the USA in
2010 [3].
The undesirable effects of fall events have led to wide
attention in the fall risk assessments, detection, and prediction
systems by health-care professionals. It is crucial for the health-
care providers to determine circumstances and scenarios that led
to a fall event and advise a mechanism to mitigate such falls [5].
In practice, most of the fall risk assessment are usually collected
through patient’s interviews and questionnaires, fall diaries,
phone calls, and simple physical performance tests [6]. Data
collection through the formerly mentioned methods provides
pertinent information, but these statistics cannot be treated
always as the reliable medical record since elderly people often
forget or fail to recall the precise circumstances of their fall event
[5], [6]. Identifying the person as a high fall-risk patient is not
sufficient to protect the patient. Therefore, accurate continuous
monitoring with a decentralized on-spot decision [7] - [9], and
recording mechanism is critical for the fall-prone patients and
elderly people [10]-[13].
Several research studies have proposed different methods to
detect a fall event, however, very few ones predicted the fall
event before it occurs [10]-[15]. The study in [6] presents an
algorithm to classify elder women subjects as high or low fall
risk patients. Ref. [15] proposes a fall prediction and detection
algorithm based on a tri-axial accelerometer, predicting the fall
event 200∼400msec before the collision with a limited number
of test cases. Falls can be detected by monitoring a person’s
surrounding using a fixed video camera with computer vision
[14]-[17], wearable cameras [18], pressure sensors [19], smart
tiles [20], and acoustic sensors (Microphone Array System) [21],

Manuscript submitted for review on October 04, 2018. This work was W. Saadeh, S. Butt, and M. A. B. Altaf are with the Electrical Engineering
funded by the Lahore University of Management Sciences (LUMS), Lahore, Department, Lahore University of Management Sciences, Lahore 54792,
Pakistan under startup grant number STG-EED-1216. Pakistan (e-mail: wala.saadeh@lums.edu.pk).

1534-4320 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2019.2911602, IEEE
Transactions on Neural Systems and Rehabilitation Engineering
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
Fig. 2: Block diagram of the proposed Fall Prediction and Detection Prototype System.
Radar-based systems [22], and Microsoft Kinetics [23]. The
alternative approach is to utilize inertial wearable sensors such
as a gyroscope and an accelerometer [24], [25] or the inertial
sensors of the mobile-phones [10], [26]-[27]. However, the
gyroscope is typically not preferable due to higher power
consumption than the accelerometer [11]. To continuously
monitor fall events, it is more suitable to use wearable devices
with detection and prediction capabilities as they can be utilized
anywhere and not limited to certain locations [24].
Several fall detection algorithms recognize multiple features
to differentiate between fall cases and regular activities of daily
living (ADL). These features include falling speed, stride time,
acceleration coordinates, posture information, inactivity periods,
and angular velocity [2], [25], [28]. The extracted features are
then classified as a fall event or an ADL, using a threshold value
or by combining posture information and kinematic thresholds,
to generate a decision [11]. Nevertheless, the falling speed and
gait features exhibit deviation from person-to-person and varies
with aging too [29],[30]. Therefore, it is essential for a fall
detection system to tackle these differences for accurate fall
detection [31]. Currently, the fall detection systems attempt to
minimize the false positives; where some ADLs are incorrectly
identified as fall events, at the cost of long latency and enhanced
computational cost [11], [32]. Moreover, current fall-detection
systems do not have a low power storage to record the fall’s
acceleration data [1].
This paper presents a patient-specific (PS) fall prediction and
detection prototype system. In this work, a single tri-axial
accelerometer sensor attached to a person’s thigh is utilized to
predict and detect the fall events and save their data for detailed
follow-up by the medical practitioner. The patient will be
notified if a fall event is predicated while the occurrence of a fall
event will initiate an alarming notice to the concern healthcare
providers by connecting it through the internet. The performance
of the proposed system is verified on 77 subjects with 600
recordings including 100 recordings of Fall/ADLs from 20 elder
subjects (aged above 65 years).
The structure of the paper is as follow. Section II describes
the proposed PS fall prediction and detection system. Section III
details the feature extraction. Section IV elaborates the proposed
fall prediction algorithm, and Section V explains the fall
1534-4320 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
detection algorithm. Section VI demonstrates the measurement
results and discussion. Finally, Section VII concludes the paper.
II. PATIENT-SPECIFIC FALL PREDICTION AND DETECTION
SYSTEM
The proposed fall prediction and detection prototype system
comprised of two parts; the data acquisition part followed by the
PS classification part. The block diagram of the proposed system
is presented in Fig. 2. In the sensing part, a tri-axial
accelerometer is used to extract the acceleration of the elderly
person in three orthogonal directions, the X, Y, and Z-axis, at a
sampling rate of 256 S/sec. The acquired accelerometer data
(Ax, Ay, and Az) along X-, Y-, and the Z-axis, respectively, are
then transferred via low energy Bluetooth interface (BLE (I/F))
to the classification part that is implemented on a Field-
Programmable Gate Array (FPGA).
A. Wearable Sensing Part
The system board is designed with an accelerometer sensor, an
Arduino (Nano) microcontroller, a 16-bit ADC, and Bluetooth
(HC-05), powered by a 9V battery with an overall size of 6 cm
× 3.5 cm × 2 cm. The sensor chosen is the MPU-6050 Tri-Axial
accelerometer with an adjustable full-scale range of ±2g, ±4g,
±8g, and ±16g, which can be attached comfortably to the
patient’s thigh without distressing the person’s routine life
activities. The upper thigh location is the preferred location for
placing the sensor since it connects the bottom kinetic chain
(legs and feet) to the parts responsible for keeping steadiness
(core and head) [6]. In addition, elder people with chances of
high fall risk show less harmonic acceleration ratio sequence in
the pelvis and upper thigh [6], [33]-[34].
The ideal wearing position of the wearable sensor is shown
in Fig. 3 (a) on the subject’s thigh. The tri-axial accelerometer
sensor is mounted inside the wearable part where the Y-axis of
the sensor is always in parallel with human body Y-axis.
However, both X-axis and Z-axis of the sensor can be
misaligned from the human body X-axis and Z-axis,
respectively, due to wearing the device in a different
angle (𝜃 ≠ 0) as depicted in Fig. 3(b). Both X-axis and Z-axis
will be rotated by the same angle θ forming two new axes X’
and Z’. The rotation will lead the sensor to read acceleration in
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2019.2911602, IEEE
Transactions on Neural Systems and Rehabilitation Engineering
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
these different axes X’ and Z’. These new axes are related to
ideal axes as given in both (1) and (2):
𝑋’ = 𝑋 𝑐𝑜𝑠(𝜃) (1)
𝑍’ = 𝑍 𝑐𝑜𝑠(𝜃) (2)
Ideally, we want 𝜃 = 0 where (𝑐𝑜𝑠 (0) = 1). For optimal
functionality of the proposed system, the rotation angle θ should
be maintained less than 20̊ (𝑐𝑜𝑠 (20) = 0.939).
Fig. 3: (a) Ideal wearing position for the wearable sensor where the sensor axes
are aligned with patient body axes, (b) Rotation of the wearable sensor forming
new axes in X’ and Z’ directions.
B. Fall Prediction and Detection System
The acquired acceleration data along each axis is processed
through a median filter followed by a high pass filter (HPF) to
mitigate the in-band noise. The HPF eliminate the gravity
acceleration since the tri-axial accelerometer generates an output
that is a combination of the desired dynamic acceleration and
gravity acceleration [25]. The system works in two parallel
modes: a fast mode for fall prediction (FMFP) and a slow mode
for fall detection (SMFD). The FMFP operates on the incoming
acceleration data at 256 S/sec while for SMFD the data are sub-
sampled to 8 S/sec. This is because, for the fall prediction, it is
required to regularly check (every 100 msec) for potential fall
event to take timely actions while for the fall detection; it is
required to look at longer period to identify a fall event (3 sec in
this system). If a large sampling rate to be used for fall detection
too, extra overhead of unnecessary computation will be
performed which will consume more power. In each operating
mode, the feature extraction block extracts the selective features
for the fall and forms a feature vector (FV). A 4KB storage
records the accelerometer data of a fall event for further
examination by physicians. If any fall event is predicted, it will
alarm the person to take precautionary actions while if a fall
event is detected, it will be transferred via the internet to the
health care providers to initiate an immediate help for the fallen
elderly person.
III. FEATURE EXTRACTION
A. Fall Prediction Features
In the FMFP, the feature extraction block extracts the
discriminating features from the tri-axial accelerometer data for
fall prediction. In order to identify the most appropriate
minimum set of features for fall prediction, several features are
examined for highest sensitivity and specificity on 600 patients’
recordings from our experiments and MobiFall dataset as
shown in Table I [6], [27]. The discriminating features for the
FMFP are carefully selected to achieve the highest sensitivity
1534-4320 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
and specificity based on features utilized in a recent study
conducted in [6]. The selected seven features provide the
highest sensitivity and specificity of 97.8% and 99.1%,
respectively. Most of these features can be also utilized for
assessing patients as high or low fall risk patients [6]. The
adopted set of features for fall prediction includes 1) mean
acceleration of X-axis(𝜇𝑥 ), 2) mean acceleration of Z-axis (𝜇𝑧 ),
3) standard deviation of acceleration of Z-axis (𝜎𝑧 ), 4)
coefficient of variance of Z-axis (𝐶𝑂𝑉𝑧 ), 5) correlation
coefficient between X-axis and Z-axis (𝐶𝑂𝑅𝑅𝑥𝑧 ), 6) mean
amplitude deviation of X-axis (𝑀𝐴𝐷𝑥 ), and 7) the total sum
vector (SV). 𝐶𝑂𝑉𝑧 is defined as the ratio between the standard
deviation 𝜎𝑧 and the mean 𝜇𝑧 :
𝜎𝑧
𝐶𝑂𝑉𝑧 = (3)
𝜇𝑧
𝐶𝑂𝑉𝑧 is utilized to detect the variability in relation to the
mean of acceleration events. 𝐶𝑂𝑅𝑅𝑥𝑧 is used to assess the
strength and direction of the relationships between X-axis and
Z-axis accelerations and is defined as follows [35]:
𝑛(∑ 𝑥. 𝑧) − (∑ 𝑥). (∑ 𝑧)
𝐶𝑂𝑅𝑅𝑥𝑧 = (4)
√( 𝑛 ∑ 𝑥 2 − (∑ 𝑥)2 ) − (𝑛 ∑ 𝑧 2 − (∑ 𝑧)2 )
Where n is the number of samples in the window, 𝑥 and 𝑧
are the filtered acceleration data in X and Z directions,
respectively.
𝑀𝐴𝐷𝑥 describes the mean distance of acceleration data points
about the mean [35]:
1
𝑀𝐴𝐷𝑥 = ∑ |𝑥𝑖 − 𝜇𝑥 | (5)
𝑛
Where n is the number of samples in the window, xi the
ith filtered acceleration sample in the X direction within the
window and µ𝑥 is the mean resultant acceleration value of the
window.
The general definition of a total sum vector (SV) magnitude
for the 3-D acceleration can be found by [25]:
𝑆𝑉 = √𝑥 2 + 𝑦 2 + 𝑧 2 (6)
Where 𝑥, 𝑦, and 𝑧 (g) denote the filtered accelerations along X-
, Y-, and Z-axis, respectively.
B. Fall Detection Features
The fall event of a person can be identified as an unintentional
sudden change from the upright/straight position to the resting
or lying position. Therefore, when a person falls, the body’s
acceleration suddenly increases. In the proposed SMFD
algorithm, only 1 feature is utilized to detect the fall, the total
sum vector square (SVS) is computed by finding the square of
the SV as given by [10]:
𝑆𝑉𝑆 = 𝑥 2 + 𝑦 2 + 𝑧 2 (7)
The proposed definition of SVS compared to SV will expand
the acceleration data range and simplifies defining the threshold
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2019.2911602, IEEE
Transactions on Neural Systems and Rehabilitation Engineering
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <

value for each subject as computing square root reduces the Receiver Operating Characteristics (ROC) of different
range of the output data [10]. machine-learning classifiers along with their Area under Curve
(AUC). ROC curves were obtained using MATLAB Statistics
Table I: List of tested features for fall prediction with corresponding sensitivity and Machine Learning Toolbox [38] by feeding 7 features for
and specificity.
pre-fall predication and looking for 3 consecutive decisions of
Features Set Sensitivity
the classifier. If all of the 3 decisions are showing pre-fall case,
Specificity
then the final output of the classifier will be also decided as a
X_MEAN, Z_MEAN, X_COV, X_SMA,
1 X_PFREQ, X_ENERGY, SV 89.3% 87.2% pre-fall case. Since the validation was based on the MobiFall
X_MEAN, Z_COV, X_SMA, X_RMS,
dataset [27], the classifier output and known outcome available
2 X_ENERGY, SV
85.5% 82.6% from the database are compared to compute the accuracy of the
Y_PFREQ, Z_COV,Y_MCR, Y_RMS,
87.0% 88.9%
system. NLSVM achieves the highest classification accuracy
3 Y_STD, Y_COV, SV compared to DT, LSVM, and KNN.
Z_COV, Z_MEAN, Z_MAD,
4 Z_STD, Z_SMA, SV
92.6% 95.4%
Z_COV, XY_CORR, X_MEAN,
5 X_MEAN, Z_STD,Z_MAD, SV
91.3% 92.6%
X_MEAN,Z_MEAN, Z_STD, Z_COV,
6 97.8% 99.1%
XZ_CORR, Z_MAD, SV
Z_COV, Z_MEAN, Z_MAD, Z_STD,
7 Z_SMA, SV
96.8% 97.5%
Z_COV, Z_MEAN, Z_MAD, Z_STD,
8 Z_SMA, SV 93.1% 96.8%
PFREQ: peak frequency SMA: signal magnitude area
STD: standard deviation MCR: mean crossing rate
CORR: correlation coefficient between two axes

IV. PROPOSED FALL PREDICATION ALGORITHM

Nonlinear Support Vector Machine Classifier (NLSVM) is
utilized in FMFP to classify the incoming features as pre-fall or
normal ADL events. Fig. 4 shows the flowchart of the NLSVM-
based fall prediction algorithm. Initially, the digitized
acceleration data will be passed to the feature extraction engine
at every 100msec to extract the discerning features. In the case
of a new patient, the NLSVM PS parameters will be weighed
by utilizing an off-line NLSVM learner [36]. Off-line learning
module based on MATLAB Statistics and Machine Learning Fig. 4: Flowchart of the fall prediction algorithm.
Toolbox [37], [38] needs a FV and former fall information for
each particular patient. The trained PS parameters are recorded
in the off-line PS repository to be utilized by on-system
NLSVM classifier. The trained parameters are then uploaded
on-processor using serial peripheral interface (SPI) protocol.
The NLSVM classifier will be ready to process the runtime FVs
after the PS parameters are uploaded in the system repository.
If the patient’s parameters are previously placed in the database,
the learning phase of the classifier will be by-passed. The
predicted fall event by NLSVM classifier will initiate an alarm
to the patient. To minimize the false alarms, three consecutive
pre-fall decisions from NLSVM classifier are required to
identify the case as a real pre-fall case and assign Fall Risk=1.
The number of pre-fall decisions for the NLSVM is optimized
Fig. 5: ROC of different machine-learning classifiers along with their
to obtain maximum prediction accuracy and maximum pre-fall achieved classification accuracy.
time. The Fall Risk alarm will be kept high for 3 sec (minimum)
such that it can be seen by the SMFD block. In the proposed The NLSVM classification operates in two stages, learning
algorithm the fall event will be predicated before (300msec- and decision phases. During the learning phase, the classifier is
700msec) of occurring as compared to 200msec-400msec in trained for a specific subject with the corresponding PS
[15]. parameters. The NLSVM classification is realized using a
The proposed NLSVM classification is selected as an Radial basis function (RBF). The RBF-NLSVM boundary
optimal choice among different machine learning algorithms, equation is defined as in (8) [36].
i.e., Linear Support Vector Machine (LSVM), Decision Tree N
(DT), and K Nearest Neighbor (KNN)). Fig. 5 shows the RBF = ∑ ( αi ∗ exp(−0.5 ∗ (‖Xi − XC ‖2 )/arg 2 )) (8)
i=1

1534-4320 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2019.2911602, IEEE
Transactions on Neural Systems and Rehabilitation Engineering
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
Where N is the number of support vectors (SVs), arg defines
the alteration of RBF, α is PS vector, and Xi and Xc are the FV
and SV, respectively, during the learning process. Xi and Xc
are both 7-dimension vectors. N, arg, α, and Xc are PS
parameters identified during the learning phase. During the
classifying phase, the real-time inward FVs (Xi) is delivered to
the NLSVM predictor with the formerly calculated PS
parameters and then the fall event is predicted based on (9) [36].
N vertical shock on the ground. The duration of this high
∑ ( αi ∗ exp(−0.5 ∗ (‖Xi − XC ‖2 )/arg 2 ))
i=1
>0 Pre − Fall
=( ) (9)
<0 No Fall
In order to compare the system performance and hardware
resources single SVS feature FMFP versus the proposed 7
features FMFP, Table II is constructed. It shows the comparison
of synthesized results based on CMOS 0.18um process of the
proposed seven-feature FV and single-feature FV i.e. SVS, with
NLSVM as prediction classifier. The single-feature FV
implementation will reduce the area and energy utilization by
38.3% and 28.1%, respectively, but at the cost of deteriorated
prediction sensitivity and specificity of 73% and 81%,
respectively, which is relatively below acceptable range for a
fall prediction system. The prediction time (latency) will remain
the same as 300-700msec due to the utilization of parallel
computation of all features in the proposed seven features
implementation.
Table II: Resources and accuracy comparison between single feature FMFP
and the proposed 7 features FMFP.
V. PROPOSED PS-THRESHOLD BASED FALL DETECTION
ALGORITHM
For distinguishing a fall event from an ADL, a fall event is
divided into three stages, i.e., the pre-fall phase, the fall impact,
and the post-fall phase. The FV formation is formed based on
tracking the SVS values that occur in 3 consecutive frames,
where 𝐸𝑖𝑗 is an indication of the 𝑆𝑉𝑆 value in each frame, ‘i’ is
the frame number 1, 2 or 3 and ‘j’ is the iteration number, shown
in (5).
𝑋𝑗 = [ 𝐸1𝑗 𝐸2𝑗 𝐸3𝑗 ] (10)
Where 𝑋𝑗 is the FV. These FVs will be used to identify the
fall events. They are explained as follows:
A. Pre-Fall Phase: A fall normally occurs when a person is
carrying out a regular movement such as walking or sitting
which can be labeled by low to medium acceleration values that
will mark the beginning of a fall case. During this phase, the 𝑆𝑉𝑆
1534-4320 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
typically have low values before the fall. Therefore, the first base
for distinguishing a fall is to have all the instances of SVS values
to be less than a predefined PS threshold (PS_TH). If this
condition is achieved, a value of ‘0’ will be stored in E1 and ‘1’
otherwise. The PS_TH needs to be updated for each patient.
B. Fall Impact: During this phase, a person will experience a
rapid falling of the body toward the ground, ending with a
acceleration spike is (300–500msec) [32]. This is shown as a
high amplitude in the acceleration curve with 𝑆𝑉𝑆 > 𝑃𝑆_𝑇𝐻.
Hence, the second base for distinguishing a fall case is a large
spike exceeding the PS_TH in the second time frame with E2
of ‘1’.
C. Post-Fall: In general, a person, after falling and hitting the
ground or any solid surface, remains in an immobile situation for
a small duration (minimum of “1-sec” time frame) before
moving again [32]. This is identified by an intermission of the
horizontal line where SVS < PS_TH which is the third base for
identifying a fall event with E3 of ‘0’.
The three above conditions represent the entire FV, which is
exploited in the proposed PS fall detection algorithm. The FV is
a 3-bit code based on the result of each frame [𝐸1 𝐸2 𝐸3 ]. Four
different configurations of frame designs are investigated to
find the ideal frame length that achieves the optimum trade-off
between detection accuracy, complexity and system latency.
Fig. 6 summarizes the comparison for different windowing
configurations. These frame designs are described as 1) a 2 sec
non-overlapped frame; 2) a 1 sec non-overlapped sliding frame;
3) a 2 sec overlapped frame with 0.5 sec augmentation, and 4)
a 1 sec overlapped frame with 0.5 sec augmentation. Analytical
results demonstrate that case 1) has the lowest hardware
complexity but with low performance (detection accuracy of
75% and the highest system latency (2 sec)). Both cases of 3)
and 4) have small latency (0.5 sec) and medium performance
(detection accuracy of 85% and 87%, respectively) but with a
huge cost for hardware implementation. Whereas, case 2)
achieves medium latency (1 sec), and high performance
(detection accuracy of >95%) with relaxed hardware
implementation cost. The FV evaluation through the non-
overlapping frame will result in reduced hardware resources
compared to the overlapping one, for example, 2) will consume
27% fewer resources compared to 4) due to a) increase in the
storage requirement and b) enhanced clock frequency for the
overlapping data. Therefore, a frame size of 1 sec is selected in
our proposed design and the FV is updated after each 1 sec with
the new incoming acceleration data. The fall event is also
detected after 1 sec from its occurrence.
Most of the conventional systems identify a fall as having
higher acceleration than ADLs that exceeds a threshold value
[25]. However, it is unreasonable to consider a constant
acceleration value as a threshold above which all motions are
considered as a fall and vice versa. Because this value may
differ for people in different situations, health conditions and
ages [32]. Selecting a low threshold may increase false
positives during ADL, resulting in lower specificity. On the
contrary, too stringent threshold selection may lead to failure of
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2019.2911602, IEEE
Transactions on Neural Systems and Rehabilitation Engineering
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <

Fig. 6: Different confiurations for sliding frame FV formation for fall detection (a) 2 sec non-overlapped frame, (b) 1 sec non-overlapped frame, (c) 2 sec
frame with 0.5 sec increment and, (d) 1 sec frame with 0.5 sec increment.

the fall detection algorithm. Fig. 7 shows the SVS for two
patients for a fall event (front fall on knees) and an ADL case
(ascending stairs). Subject A is 30 years old male with a height
of 177 cm and weight of 102 Kg, while Subject B is a 26 years
old female with a height of 170 cm and weight of 90 Kg. From
Fig. 7 (a) and (c), Subject A needs to have a high PS_TH to
detect fall case correctly (~1.0), while Subject B in Fig. 7 (b)
and (d) needs a smaller value of PS_TH (~0.75). Selecting the
same threshold value for both subjects will erroneously detect
an ADL as a fall or vice versa. The PS_TH value is determined
once for each patient off-chip. The PS threshold (PS_TH) is
computed off-line with the help of MATLAB and uploaded
through SPI protocol to the FPGA. In this work, linear
regression is utilized to determine the optimal PS_TH to
perform on-sensor processing and reduce the Bluetooth data
overhead. Each subject’s PS_TH need to be determined,
therefore, the training set is required.
The flow chart of the proposed fall detection algorithm is
presented in Fig.8. Fall detection decision formation is based on
the 3-bits FV formed by the 3 consecutive “1-sec” frames. A fall
event is identified by a ‘010’ code as it resembles a high
Fig. 7: SVS for two subjects for a fall and no fall cases (a) and (b) are the
amplitude in the central frame only. For the cases where the last fall cases for patients A and B, respectively, (c) and (d) are the ADL case for
two bits are ‘01’, they can be fall events. Hence, the verdict is patients A and B, respectively.
delayed until the next cycle after receiving the adjacent 1-sec
time frame of the acceleration data. In rare cases where a fall
VI. MEASUREMENT RESULTS AND DISCUSSION
event produces two spikes larger than the PS_TH value in two
consecutive 1-sec frames, the SMFD system will check if there To verify the working of the proposed system in the real
was a fall event predicated by FMFP to decide a fall or no fall environment, it was tested on elder subjects. Despite the non-
case. The coordination between the two modes of the systems invasiveness nature of the proposed system, the Institutional
will make the overall system more reliable and reduce false Review Board (IRB) approval was processed along with the
negatives. The remaining cases are considered ADLs since the participant consent form under the IRB Organization number
conditions for fall event are not met. If a fall event is identified, LUMS-IRB20170114 for the device testing in human subjects.
the tri-axial acceleration information is kept in the local All the falls and ADLs are performed in a realistic way similar
repository and an alarming signal is transmitted to the health to the method utilized in MobiFall dataset [27]. However, we
care provider through the internet. The fall will be spotted after used a thicker mattress (10 cm) to dampen the fall and protect
1 sec of the occurrence which is >50% less latency than the elder subjects who participated in the experiments. The
[1],[2],[23]. Fall/ADL validation experiments are performed by 20 different
subjects (12 males and 8 females) (age: 65-70) (Body Mass

1534-4320 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2019.2911602, IEEE
Transactions on Neural Systems and Rehabilitation Engineering
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <

Index (BMI): 28.3-35.1). The validation experiments combined
with the MobiFall dataset comprise of 6 different fall cases and
11 multiple ADLs from a total of 77 subjects with greater than
600 recorded trials. The subjects in the MobiFall dataset are
healthy 42 males (age: 20-47 years, BMI: 19.78-34.68) and 15
females (age: 20- 36 years, BMI: 18.36-31.14) [27]. The ADLs
are selected based on their consistency and resemblance to
realistic falls that may increase the false positives. Explicitly,
they are chosen established on the subsequent norms: a) fall-like
events where the subject typically stays immobile at the end, for
example, stepping in a car or sitting on a seat; b) ADL like
standing, walking, and descending and ascending the stairs; c)
abrupt and rapid events that are alike to fall events, such as
jogging and jumping. The four falls comprise a) forward lying
fall, b) back chair fall, c) front knees fall, and d) side fall.
Fig. 10 (a) shows the logic resources utilized by the proposed
SMFD/FMFP system (feature extraction and classification) on
FPGA Virtex5 (XC5VLX110T), which result into equivalent
logic cells of 62.3K [39]. Power budget distribution of the
sensor, microcontroller, and FPGA for the proposed system is
shown in Fig. 10 (b). The power distribution chart focuses
mainly on processing power with SMFD/FMFP consumes
0.34mW at an operating clock of 1 KHz, and does not include
wireless transmission power, which will dominate the power
budget.

Fig.9: Application of different scenarios for fall detection algorithm on one

subject (male, 65 years, and BMI of 28.3) (a) Jumping, (b) Sitting on a chair,
(c) Fall on the right side, and (d) Forward from standing, use of hands to dampen
fall.

Fig.8: Flowchart of the proposed fall detection algorithm.

Fig.9 presents the measurement results of one healthy subject
(Male, 65 years, BMI of 28.3) performing four distinct tasks
(jumping, sitting on the seat, falling on the right side, and falling
on the front side). The fall prediction features are evaluated
every 100msec (not shown in Fig. 9) and fed into NLSVM
classifier for pre-fall decision as described in Eq. (9). If a pre-
fall event is detected, a Fall Risk alarm will be set to 1 for 3 sec.
The SVS feature for fall detection is shown in Fig. 9. For both
cases of ADL in Fig.9 (a) and (b), the fall risk remains as 0 while
performing these normal daily activities which means that no fall
is predicted. However, in the cases of fall events as shown in
Fig.9 (c) and (d), the fall event is predicted as fall risk alarm
Fig.10: a) FPGA resource utilization for the proposed SMFD/FMFP system, b)
Power distribution of the overall system.
To compare the performance of the proposed FMFP and
SMFD algorithms confusions matrix are formed as shown in
Fig. 11. The evaluation shows that, for 600 tests including
measured experiments and MobiFall dataset results, the
proposed FMFP algorithm accomplishes sensitivity and
specificity of 97.8% and 99.1%, respectively, with maximum
prediction time of 700 msec before the fall impact.
Fall Detection Fall Prediction
True Class

99.1%
True Class

raised to 1 with pre-fall impact time of 400msec and 600msec, 99.3% 0 99.1% 0.9%
0 99.3% 0.7% 0.9%
respectively. From the offline training of the proposed SMFD 0.7%
system, PS_TH is computed to be 1.5 for this subject. Each case 98.6% 1 2.2% 97.8% 97.8%
1 1.4% 98.6%
1.4% 2.2%
in Fig.9 resulted in a unique 3-bit output FV that is passed to the 0 1
0 1
detection algorithm for fall detection. The proposed SMFD TPR/FNR Predicted Class TPR/FNR
algorithm correctly identifies jumping and sitting on a seat as No
(a) (b)
Fall and front knee- fall and falling on right sideways as fall
Fig. 11: Confusion Matrix of the proposed fall a) detection and b) prediction.
cases.

1534-4320 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2019.2911602, IEEE
Transactions on Neural Systems and Rehabilitation Engineering
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
Table III. Comparison between proposed algorithm and state-of-art work.
The proposed SMFD algorithm achieves sensitivity and
specificity of 98.6% and 99.3%, respectively. It achieves a small
latency of 1 sec for fall detection, which is >50% less time than
[1], [2], [23]. It achieves superior performance than the multi-
sensor [2] and multi-threshold [25] acquisition and algorithms,
respectively. Table III compares the proposed algorithms with
the state-of-the-art works on fall-prediction and detection. Ref
[15] achieves slight better sensitivity and specificity compared
to the proposed FMFP but does not report latency of the system
and utilizes the Hidden Markov Model which will be hardware
costly compared to the proposed implementation. Moreover, the
reported prediction accuracy is based on 8 young people’s
simulated activities and the accuracy will be severely affected if
tested in a real-time environment due to the utilization of non-
PS approach. Offline training for NLSVM parameters and PS-
threshold findings reduces the hardware complexity and high
power utilization burden on the wearable device system.
VII. CONCLUSION
In this work, we propose a PS single sensor fall prediction and
detection prototype system. It will alarm the patient to take
action if a fall case is predicted while if a fall event is detected,
it will be communicated to health care providers. The detected
fall events will be alarmed to the health care providers through
Clouds to provide immediate help to the fallen elder person.
The accuracy of the proposed algorithms is validated via
MobiFall Dataset and our performed experiments. FMFP
achieves sensitivity and specificity of 97.8% and 99.1%,
respectively, while SMFD achieves sensitivity and specificity
of 98.6% and 99.3%, respectively, for a total 600 measured
ADL and falls from 77 subjects. The performance can be
improved by studying further falls and fall-like events.
1534-4320 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
REFERENCES
[1] D. Chen, et al., “A wearable wireless fall detection system with
accelerators,” in Proc. IEEE Int. Conf. Robot. Biomimet. (ROBIO), Apr.
2011, pp. 2259-2263.
[2] W. Baek, et al., “Real life applicable fall detection system based on
wireless body area network” in Proc. IEEE Consum. Commun. Network.
Conf. (CCNC), Jan. 2013, pp. 62–67.
[3] Injury Prevention and Control: Data and Statistics (WISQARSTM),
[Online]. Available: http://www.cdc.gov/injury/wisqars/
[4] “Falls Prevention Facts”, [Online] Available,
https://www.ncoa.org/news/resources-for-reporters/get-the-facts/falls-
prevention-facts/
[5] Y.S. Delahoz, et al., “Survey on Fall Detection and Fall Prevention Using
Wearable and External Sensors,” Sensors, vol. 14, no. 10, pp. 19806–
19842, Oct. 2014.
[6] A. Hua, et al., “Accelerometer-based predictive models of fall risk in
older women: a pilot study,” Nature Digital Medicine, 2018.
[7] S. Abubakar, et al., “A Wearable Auto-Patient Adaptive ECG Processor
for Shockable Cardiac Arrhythmia,” IEEE Asian Solid-State Circuits
Conf. (ASSCC), Nov. 2018, pp. 267-268.
[8] W. Saadeh, et al., “A 1.1mW Ground Effect-Resilient Body-Coupled
Communication Transceiver with Pseudo OFDM for Head and Body Area
Network,” IEEE Journal of Solid-State Circuits (JSSC), vol. 52, no. 10,
pp. 2690- 2702, Oct. 2017.
[9] S. Abubakar et al. “A Wearable Long-Term Single-Lead ECG Processor
for Early Detection of Cardiac Arrhythmia,” Proc. IEEE Design
Automation and Test in Europe (DATE), pp. 961-966 Mar. 2018.
[10] W. Saadeh, et al., “A High Accuracy and Low Latency Patient-Specific
Wearable Fall Detection System,” IEEE Biomedical and Health Informa
tics (BHI), 2017, pp. 441-444.
[11] S. Abbate, et al., “Recognition of false alarms in fall detection systems,”
in Proc. IEEE Int. Work. Consum. eHealth Platforms, Services, and
Applications, CCNC Workshop, Jan. 2011, pp. 538–543.
[12] M. Altaf, et al., “A 0.21µJ Patient-Specific REM/Non-REM Sleep
Classifier for Alzheimer Patients,” in Proc. IEEE Biomedical Circuits and
Systems (BioCAS), Oct. 2017, pp. 652-655.
[13] P. Rashidi and A. Mihailidis, “A survey on ambient-assisted living tools
for older adults,” IEEE J. Biomed. Health Informat., vol. 17, no. 3, pp.
579–590, May 2013.
[14] B. Mirmahboub, et al., “Automatic monocular system for human fall
detection based on variations in silhouette area,” IEEE Trans. Biomed.
Eng., vol. 60, no. 2, pp. 427–436, Feb. 2013.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2019.2911602, IEEE
Transactions on Neural Systems and Rehabilitation Engineering
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <

[15] L.Tong, et al., “HMM-Based Human Fall Detection and Prediction

Method Using Tri-Axial Accelerometer,” IEEE Sensors Journal, vol. 13,
no. 5, pp. 1849-1856, May 2013.
[16] E. Auvinet, et al., “Fall detection with multiple cameras: An occlusion-
resistant method based on 3-D silhouette vertical distribution,” IEEE
Trans. Inf. Technol. Biomed., vol. 15, no. 2, pp. 290–300, Mar. 2011.
[17] C. Rougier, J. Meunier, A. St-Arnaud, and J. Rousseau, “Robust video
surveillance for fall detection based on human shape deformation,” IEEE
Trans. Circuits Syst. Video Technol., vol. 21, no. 5, pp. 611–622, May
2011.
[18] K. Ozcan, et al., “Autonomous Fall Detection With Wearable Cameras by
Using Relative Entropy Distance Measure,” IEEE Trans. Human-
Machine Interface sys. vol. 47, NO. 1, pp. 31-39, Feb. 2017
[19] H. Rimminen, et al., “Detection of falls among the elderly by a floor
sensor using the electric near field,” IEEE Trans. Inf. Technol. Biomed.,
vol. 14, no. 6, pp. 1475–1476, Nov. 2010.
[20] M. Daher, et al., “Elder Tracking and Fall Detection System Using Smart
Tiles,” IEEE Sensors Journal, vol. 17, no. 2, pp. 469-479, Jan. 2017
[21] Y. Li, et al., “A microphone array system for automatic fall detection,”
IEEE Trans. Biomed. Eng., vol. 59, no. 5, pp. 1291–1301, May 2012.
[22] C. Garripoli, et al., “Embedded DSP-Based Telehealth Radar System for
Remote In-Door Fall Detection,” IEEE J. Bio. Health Inform. (JBHI), vol.
19, no. 1, pp. 92-101, Jan. 2015.
[23] E. E. Stone, et al., “Fall Detection in Homes of Older Adults Using the
Microsoft Kinect,” IEEE J. Bio. Health Inform. (JBHI), vol. 19, no. 1, pp.
290-301, Jan. 2015.
[24] Q. Li, et al., “Accurate, Fast Fall Detection Using Gyroscopes and
Accelerometer-Derived Posture Information,” IEEE Int. Work. Wear.
Implant. Body Sensor Networks, Jun. 2009, pp. 138-143.
[25] A. Sorvala, et al., “A two-threshold fall detection algorithm for reducing
false alarms,” IEEE Int. Symp. Med. Info. Commun. Techn. (ISMICT),
May 2012, pp. 1-4
[26] L.-J. Kau, and C. Chen, “A Smart Phone-Based Pocket Fall Accident
Detection, Positioning, and Rescue System,” IEEE J. Bio. Health Inform.
(JBHI), vol. 19, no. 1, pp. 44-56, Jan. 2015,
[27] G. Vavoulas, et al., “The MobiFall Dataset: Fall Detection and
Classification with a Smartphone”, Int. J. Monit. Surveill. Techn.
Research, pp 44-56, 2014.
[28] W. Saadeh, et al., “A Wearable Neuro-Degenerative Diseases Classifier
System based on Gait Dynamics,” IEEE VLSI-SoC, Oct. 2017, pp. 1-6.
[29] S. Ko, et al., “Age-Associated Differences in the Gait Pattern Changes of
Older Adults during Fast-Speed and Fatigue Conditions: Results from the
Baltimore Longitudinal Study of Ageing,” Age and Ageing, vol. 39, no.6,
pp. 688–694, 2010.
[30] R. Bartlett, “Introduction to Sports Biomechanics: Analyzing Human
Movement Patterns”, 2nd Edition, Routledge, 2007
[31] B. J. Fregly, et al., “Design of patient-specific gait modifications for knee
osteoarthritis rehabilitation,” IEEE Transactions on Biomedical
Engineering, vol. 54, no. 9, pp. 1687-1695, Sept. 2007.
[32] M. Yu, et al., “An online one class support vector machine-based person-
specific fall detection system for monitoring an elderly individual in a
room environment,” IEEE J. Biomed. Health Informatics, vol. 17, no. 6,
pp. 1002–1014, Nov. 2013.
[33] H. B Menz, et al., “Acceleration patterns of the head and pelvis when
walking are associated with risk of falling in community-dwelling
older people,” J. Gerontol. A. Biol. Sci. Med. Sci., vol. 58, no. 5, pp.
446–452, 2003.
[34] N. Noury, et al., “A proposal for the classification and evaluation of fall
detectors,” IRBM, vol.29, pp. 340–349, 2008.
[35] R. Wilcox “Modern Statistics for the Social and Behavioral Sciences: A
Practical Introduction”, CRC Press, 2011.
[36] M. Altaf and J. Yoo, “A 1.83μJ/Classification, 8-Channel Patient-Specific
Epileptic Seizure Classification SoC using Non-Linear Support Vector
Machine,” IEEE Trans. Biomed. Circ. Syst. (TBioCAS), vol. 10, no. 1, pp.
49-60, Feb. 2016.
[37] V. Franc and V. Hlavac, Statistical Pattern Recognition Toolbox for
Matlab, Czech Technical University, 2004, pp. 27–33.
[38] Matlab Statistics and Machine Learning Toolbox, [Online] Available:
https://www.mathworks.com/help/stats/
[39] Virtex-5 FPGA User Guide, [Online] Available:
https://www.xilinx.com/support/documentation/user_guides/ug190.pdf

1534-4320 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.